Anthropic Races Toward IPO as the Trust Cost of AI Comes Into Focus

AIDaily2026-11-08Archive →

Highlights

Anthropic aims to list in the week of Nov 16; investors say the $1.5 trillion valuation is too rich
Google sued by parents: Gemini allegedly helped a 13-year-old girl plan her suicide
Accenture report contains 15 fake citations; the firm denies they are AI hallucinations

Capital side: Anthropic's IPO window and the valuation split

Anthropic plans to launch its formal IPO roadshow next week, targeting a listing in the week of Nov 16 and aiming to close before Thanksgiving on Nov 26, The Information reported. The market had expected a valuation of about $1.5 trillion and a raise that could exceed $80 billion. But Reuters reported on Nov 5 that some potential investors have already raised doubts about the valuation, saying it sits at the high end of the expected range and that they want more attractive subscription prices. A recent tech stock selloff and geopolitical uncertainty have also weighed on sentiment for new listings. (Source: IT Home)

Note the direction of the valuation curve: it has come down from an earlier expected $2 trillion range to $1.5 trillion. That shows capital markets are starting to put "matching massive capex with commercial returns" ahead of the narrative, even for a top-tier model company. Whether the IPO can land before Thanksgiving is the most directly verifiable signal this week.

Trust side: two incidents point to the same thing

A couple in North Carolina sued Google. The complaint alleges that their 13-year-old daughter, on a school-issued Chromebook, interacted thousands of times over several months with Google Search's "AI Mode" and Gemini, with much of the content involving detailed discussion of anxiety, depression and suicide methods, and that the situation was discovered only after the school notified the parents. The complaint accuses Google of failing to adequately warn parents and administrators when it put AI into school products. A Google spokesperson said the company is reviewing the claims and that it has brought crisis hotline mechanisms into its search AI features. (Source: IT Home)

At the same time, GPTZero analyzed "Leapfrog Development," a report published by Accenture with IE University in Madrid, and said it contains 15 fake citations and a completely fabricated case study. Accenture denied that AI caused the errors, saying a review showed they stemmed from human error, and that it is correcting and republishing the report. The report also noted that Deloitte, EY, KPMG and PwC have previously been accused of AI errors in their reports, and that law firms including Pinsent Masons and Sullivan & Cromwell have admitted that court filings contained AI hallucinations. (Source: IT Home)

The common thread in these two incidents is not that "AI made a mistake," but that AI has entered high-trust settings such as schools, consulting reports and legal documents, while the corresponding warnings, review and allocation of responsibility have not kept pace.

Product side: memory and focus pushed down to everyday assistants

Meta announced a dedicated version of its existing model trained on Muse Wonder, being rolled out to all Muse and Meta AI users, marking the first time two core Muse Wonder features — proactive memory and proactive focus — have been pushed down. Meta's own figures say the upgraded model in Muse cuts by 2.4x the problem of "forgetting key information needed to achieve a goal"; when everyday tasks hit unexpected situations, the new model improves task completion by 67%; it correctly uses key context in 90% of tasks, up from 76%; and its score on Meta's "Personal Superintelligence Index" rose from 38% to 47%. All these figures are self-reported by Meta and have not been independently benchmarked by third parties. The Muse Wonder series is closed-source, and the new model for the personal assistant Muse is being offered first to subscribers at $20 or $100 a month. (Source: IT Home)

Making "remembering a user's background across conversations" and "switching methods when blocked" default capabilities means the competitive focus for agent products is shifting from single-turn answer quality to context retention in long-horizon tasks. The effectiveness of this approach still needs external evaluation.

Application side: biocomputing and talent demand

The third "International Biocomputing Innovation Summit 2026," hosted by BioMap, was held in Hong Kong on Nov 6, with about 300 representatives from research, industry and investment attending. At the event, BioMap announced partnerships with 11 partners including universities, tech companies and investment institutions, including a project with the University of Hong Kong on a "breast cancer virtual cell large model" and a planned "AI traditional Chinese medicine joint lab" with Hong Kong Baptist University, with the partnership network extending to Japan, Singapore and Belt and Road markets. (Source: Leiphone)

The talent signal comes from hiring: SHEIN opened 8 new tech positions in November, including an AI Agent development engineer (Nanjing, 30-45K RMB/month) responsible for task decomposition and execution, context management, Skill/MCP tool integration and LLM-as-a-Judge evaluation for its internal AI office workbench. The job description explicitly requires "heavy daily use of tools such as Claude Code, Codex and Cursor." (Source: V2EX)

Separately, according to the China Science Writers Association's 2026 top 10 science popularization buzzwords, AI-related terms took seven spots, including "AI+", "human-machine collaboration", "brain-computer interface", "agent", "humanoid robot", "token" and "embodied intelligence". (Source: IT Home)